[NeurIPS 2024] SteerEval: Mapping the Boundaries of LLM Controllability
How Controllable Are Large Language Models? A Unified Evaluation across Behavioral Granularities
The paper introduces SteerEval, a hierarchical benchmark for evaluating the controllability of Large Language Models (LLMs) across three domains: personality, sentiment, and language features. It structures evaluation into three granularity levels—L1 (intent), L2 (strategy), and L3 (instantiation)—providing a systematic framework to test contemporary steering methods like RePS, DiffMean, and PCA.
TL;DR
Can we truly control how an LLM behaves? While we can easily nudge a model to be "happy," forcing it to follow specific stylistic strategies or strict token-level constraints is a different beast entirely. SteerEval is a new hierarchical benchmark that reveals a sobering truth: current steering methods (like Activation Steering) fail precisely where they are needed most—at the fine-grained level of implementation.
The Granularity Gap: Why "Be Nicer" Isn't Enough
Most research in LLM "steering" focuses on broad strokes—changing the sentiment from negative to positive or shifting a persona. However, the authors of SteerEval argue that controllability exists on a spectrum of granularity.
Imagine you want a model to be Autonomous (Personality Domain):
- Level 1 (Computational): The model should just "express autonomy."
- Level 2 (Algorithmic): The model should "use self-directed choices."
- Level 3 (Implementational): The model must use the specific word "self-authored."
Current benchmarks often conflate these, but SteerEval separates them to find exactly where our steering wheel loses its grip.
Methodology: Marr's Vision for LLMs
The authors leverage a classic framework from cognitive science: Marr's Three Levels of Analysis. They map this to LLM behavior to see if steering vectors can influence the model's internal hierarchy.
The Synthesis Pipeline
To build the benchmark, the team used an automated pipeline to:
- Synthesize Hierarchies: Generate 8 independent concepts per level across Personality, Sentiment, and Language Features.
- Question Rewriting: Pivot questions toward a related-but-distinct concept to avoid "shortcut" cues in the prompt.
- Paired Answers: Create (Matching, Not-Matching) answer pairs with minimal lexical edits to isolate the concept being tested.
Experimental Showdown: Prompting vs. Activation Steering
The researchers tested three major LLM families (Gemma, Qwen, Llama) using two main steering paradigms:
- Prompt-based: 0-shot and 3-shot prompting.
- Activation-based: PCA, DiffMean, and RePS (Representation Steering).
Key Findings: The "Decline" of Activation Steering
The results were striking. While Prompting remained consistently strong across all levels (L1 through L3), Activation Steering methods—previously thought to be the "cleaner" way to control models—collapsed as the constraints became more specific.
As seen in the data above, on Gemma-2-9B, the Harmonic Mean (HM) for activation methods plummeted. For example, PCA dropped from a modest 1.85 (L1) to a near-zero 0.15 (L3). This suggests that while we can find a "direction" for sentiment in the hidden states, finding a direction for "use exactly this word while maintaining logic" is far more elusive.
Visualizing the Failure: A Case Study
The paper provides a qualitative look at how different levels impact output. In Level 1, the model easily adopts a redundant style. By Level 3, the model is asked to use a specific marker (like "i.e."). Activation steering often fails here, either ignoring the marker or becoming completely incoherent (low fluency).
Critical Insight: Why Does This Happen?
The authors hypothesize that different domains occupy different depths in an LLM's internal hierarchy.
- Personality is a high-level "dispositional prior."
- Language Features are low-level "surface realizations."
Activation steering transfers less uniformly across these domains. Furthermore, increasing "steering strength" often improves the Concept Score but destroys Instruction Following and Fluency. We are essentially "overclocking" the neurons for one feature at the expense of the model's general intelligence.
Conclusion & Future Outlook
SteerEval provides a much-needed reality check for the AI Safety and Alignment communities. If we cannot reliably steer a model to follow a simple "L3" lexical constraint using internal vectors, how can we expect to steer it toward complex, abstract ethical principles using the same methods?
The Road Ahead: Future research needs to bridge the gap between prompting's robustness and activation steering's interpretability. We need "surgical" interventions that can handle the L3 granularity without breaking the model's ability to speak.
Senior Editor's Note: This paper highlights the "illusion of control" in current LLM steerability research. While we have mastered the 'vibes,' we are still far from mastering the 'mechanics'.